Impact of partial volume correction on radiomics reproducibility in theranostic SPECT/CT imaging

Abstract

Background: Radiomics has shown potential for quantitative characterization of tumors in molecular imaging; however, its clinical translation in theranostic ¹⁷⁷Lu SPECT/CT remains limited due to poor robustness of extracted features to reconstruction variability and partial volume effects. Establishing reproducible radiomics biomarkers across correction strategies is therefore a prerequisite for reliable clinical modeling and treatment monitoring.

Purpose: This study aimed to evaluate radiomics feature reproducibility, defined as the stability of feature values across different partial volume correction (PVC) strategies and reconstruction settings, in clinical ¹⁷⁷Lu SPECT/CT imaging. In addition, we explored two volumetric shape-based indices, the metastasis-to-liver ratio (MLR) and metastasis-to-spare liver ratio (MSLR), as surrogate markers of hepatic metastatic burden in the theranostic treatment setting.

Methods: In 13 patients (40 scans) treated with ¹⁷⁷Lu, 837 radiomics features were extracted from 11 abdominal regions and metastases on SPECT/CT using original and wavelet-decomposed images across four bin widths (50–200). Two post-reconstruction PVC methods, namely Richardson-Lucy (RL) and Reblurred Van Cittert (RVC), were applied. Feature reproducibility was quantified using two complementary metrics: the intraclass correlation coefficient (ICC) to assess feature-level stability across PVC strategies, and the concordance correlation coefficient (CCC) to evaluate pairwise agreement and systematic bias among reconstruction methods. Visual image quality assessments were independently performed by two experienced nuclear medicine specialists in a blinded setting. Exploratory metastatic tumor burden was assessed descriptively using 3D shape-based MLR and MSLR indices.

Results: Low-frequency wavelet decomposition (LLL-wavelet) and original features showed the highest reproducibility (ICC ≥ 0.90 in >95% of liver and metastasis features at BW50), whereas high-frequency features and larger bin widths demonstrated reduced stability. CCC analysis revealed excellent agreement between RL and RVC (≥0.95 in major organs at BW50–100), while agreement with uncorrected SPECT (no PVC) was consistently lower, especially for high-frequency features. RL achieved higher visual scores in sharpness and contrast (p < 0.01), with good inter-reader agreement supporting the consistency of these assessments. MLR/MSLR demonstrated inter-patient variability and were explored descriptively as indices of metastatic liver burden.

Conclusions: Reproducibility in theranostic SPECT radiomics is highly feature- and organ-dependent and is further influenced by scanner-specific factors and reconstruction protocols, which remain critical for real-world clinical translation. RL and RVC showed stronger mutual agreements than each with uncorrected SPECT. Importantly, only RVC translated visual improvements into enhanced feature-level reproducibility, while RL provided the most consistent overall balance of reproducibility and image quality, supporting its role as the preferred PVC strategy for clinical and modeling applications. Robust radiomics feature selection as well as standardized reproducible PVC strategies are essential to generate methodological harmonization for future clinical translation and to support integration of radiomics analyses into personalized SPECT theranostics.